Building Condition Models using Falkonry Patterns¶
Building a multivariate model using Falkonry Patterns involves defining the use case, setting up datastreams and signal groups, labeling events, training an unsupervised or semi-supervised model, and refining those models until objectives are met.
Create a Datastream¶
Datastreams are core to organizing your model. Each datastream typically corresponds to a single equipment or entity with fewer than 15 signals.
- Create a Datastream by importing a source manifest (CSV or Parquet)
- Add Signal Properties such as sendOnChange, sampleInterval, minThreshold, and maxThreshold
Signal Groups are subsets of signals for modeling or reference.
- Use prefixes like _model-signals or _reference-signals for clarity
- Keep modeling groups under 20 signals
- A single signal may belong to multiple groups used in modeling
Example of a Datastream Manifest¶
| source_name | datastream_name | entity_name | signal_name | signal_description | sendOnChange | gapThreshold |
|---|---|---|---|---|---|---|
| machine1/current1 | Machine monitoring | machine1 | current1 | output current reading | true | PT2H |
| machine1/current2 | Machine monitoring | machine1 | current2 | outer current reading | true | PT2H |
| machine3/pressure1 | Machine monitoring | machine3 | pressure1 | outer casing pressure | true | PT2H |
| machine2/temperature1 | Machine monitoring | machine2 | temperature1 | outer casing pressure | true | PT2H |
| ... | ... | ... | ... | ... | ... | ... |
Defining Events in Falkonry Patterns¶
Click here for details about Event Manager in Falkonry Patterns.
Define Events (Facts)¶
Events represent labeled time periods and are used to train supervised or semi-supervised models.
- Identify 3–5 time periods each for normal, warning, failure, shutdown, etc.
- Avoid using failure events directly for early warning models
- Create/import events via the Timeline or CSV (start, end, entity, event columns)
- For Event Groups, use prefixes like _normal, _warning, etc., and keep facts focused on central portions of each condition
Best Practices:¶
- Include more than one label type for multi-class classification
- Keep labels concise and non-overlapping
- Avoid broad or transitional periods that dilute condition clarity
Example of an Events CSV File¶
| time | end | entity | value | event groups |
|---|---|---|---|---|
| 2024-08-31T19:50:00.000Z | 2024-09-13T20:15:00.000Z | machine1 | training | _training |
| 2024-09-01T03:11:17.340Z | 2024-09-03T18:08:11.725Z | machine1 | normal | _supervision |
| 2024-09-04T05:08:00.000Z | 2024-09-04T21:36:00.000Z | machine1 | warning | _supervision |
| 2024-09-04T21:56:00.000Z | 2024-09-05T02:24:00.000Z | machine1 | fault | _supervision |
| 2024-09-05T02:44:00.000Z | 2024-09-06T00:00:00.000Z | machine1 | eval. | _evaluation |
Building a Unsupervised Model¶
Building a Semi-supervised Model¶
TSI Clip Coming Soon
Building a Supervised Model¶
TSI Clip Coming Soon
Click here for more details Modeling Workflow for building a Condition model.